Native MiniMax H3 inference with explicit GPU profiles. Vflash uses PyTorch and Triton, with its own denoising runtime and support for pinned LightX2V Turbo LoRAs.
Documentation · Get started · Release notes · 中文
0.22.1: native HTTP response contracts. OpenAPI now describes readiness, job lifecycle, errors and binary/range downloads; inference is unchanged. See HTTP API.
0.22.0: segmented motion, texture and native sound. The new explicit SM120 reference runtime defaults to Veda with four Larry plus six original visual evaluations, followed by twelve Larry native-audio evaluations. See scope and asset contract.
0.21.0: combine appearance images and motion video. The hybrid pipeline accepts ordered image references alongside one video. Explicit video-only spatial pooling preserves the appearance latents and lowers motion-context cost. Initial complete clips ran on RTX 4090 48 GiB and RTX PRO 6000 Blackwell Server; see hardware timings and measured scope.
0.17.2: choose the Veda layer partition independently of precision. Set
veda_dense_edge_layers=4 for 42 sparse / 8 dense layers, including the optional
audio context pass. SM120 keeps the selected Sage FP8 or FROST precision policy.
The H3Pipeline Veda partition default stays 40/10 (the explicit segmented runtime uses 44/6); see matched execution results.
0.16.0: faster SM120 attention. On RTX PRO 6000, veda_dense_backend="sageattention2" now runs
INT8 Q·K / FP8 P·V in both dense and Veda sparse layers: denoising 6–14% shorter, with flicker and
sharpness measured against the BF16 baseline. See the release notes.
0.15.0: resident 4 MP on 96 GiB Blackwell. Resident I2VA on one SM120 GPU now follows the native-HD pixel-frame budget instead of a fixed 1536×864 canvas: an RTX PRO 6000 with the FP8 trunk, face and six-step audio repair completed 1920×1088/5 s in 163 s, 2720×1536/5 s in 261 s and 1920×1088/10 s in 245 s at a 42.1 GiB peak. See residency.
0.14.0: true 4 MP native HD. Image-conditioned native HD keeps the 2048² pixel cap but accepts a 2720-pixel long edge, so five-second requests reach 2720×1536 (16:9), 1536×2720 (9:16) or 2368×1760 (4:3). Five complete RTX 4090 requests across these shapes stayed coherent; the joint pixel-frame budget is unchanged. See the canvas limits.
0.13.0: opt-in audio hidden-context reuse. The owned six-step audio interface can freeze first-evaluation non-audio hidden states, with current audio and rebuilt K/V on every later step. Two 512²/5s comparisons on a 48 GiB RTX 4090 retained every video frame and received listening acceptance, with about 34% less audio time. This is an explicit approximation, disabled by default; larger cached workloads remain unqualified. See the audio guide.
The optional six-step audio pass supports the native BF16 weight ring and qualified resident FP8 paths with Veda 44/6 and approximately 0.5 MP spatial context. Complete video pixels remain unchanged. The pipeline guide separates hardware qualification, listening evidence and execution cost.
Vflash provides two interfaces:
- Complete video generation:
H3Pipelineandvflash generateown conditioning, native denoising, official video/audio decoding and MP4 encoding. Text, keyframes and references use distinct prepared profiles. Original LightX v0.1 and hybrid profiles support the qualified duration, reference and pixel-frame limits. - Native inference service: a resident GPU process accepts compiled conditioning bundles and returns video/audio latents. The bounded HTTP queue owns no user accounts or durable jobs; the integrating application owns authentication, scheduling and media delivery.
Explicit profiles cover single SM89, cooperating SM86/SM89 pairs, and qualified H100 and Blackwell configurations. Exact GPU memory, mode and geometry matter; consult the current profile guide rather than treating hardware names or a historical release as a capacity guarantee. No automatic device substitution or attention fallback occurs.
Reusable video repaint and face-encoding APIs are opt-in building blocks. Face detection, track selection, restoration models and final compositing belong to the integrating application. Installing the engine alone does not enable a product's full face-restoration workflow. Existing model, precision and attention defaults remain unchanged.
Python 3.11 or newer is required. The base install does not download model weights or PyTorch.
git clone --branch v0.19.0 --depth 1 https://github.com/Hansimov/vflash.git
cd vflash
python -m venv .venv
source .venv/bin/activate
python -m pip install -e .
vflash doctor
vflash profiles
vflash plan ref2va-turbo4-exact-sm89 --gpu 0For cooperating GPUs, select the peer explicitly with --peer-gpu; one process owns the entire
group until it closes. Host memory must cover weights, encoders, decoders and the selected request.
See hardware and memory limits.
- Generate an MP4 with the complete Ref4 or T2VA pipeline and its pinned official encoder/VAE adapters.
- Prepare official weights, or run a bundle with existing compatible assets.
- Integrate through Python and reuse a loaded model across requests.
- Start Docker and HTTP for an isolated worker and bounded job queue.
- Measure speed and quality, with loading and end-to-end costs kept separate.
- Compare Sol-Engine mechanisms against the actual SM86/SM89 contract.
- Troubleshoot startup, memory, dual-GPU and output problems without weakening ownership checks.
Turbo4 and Turbo8 are distilled adapters. Exact attention is not a base-model quality guarantee, and different GPU or parallel configurations need not produce bitwise-identical results. Complete generation uses explicit fixed-profile assets. Dual-SM86 Ref4 requires sequence-head; single-SM86 and other native-only Turbo profiles retain their latent interface. W8 and arbitrary adapter or mode switching are outside the supported profiles. Read the validation scope.
python -m pip install -e '.[dev,server]'
pre-commit install
pre-commit run --all-files
pytestBuild the bilingual docs with npm ci and npm run docs:build. See the
contribution guide and contributor map
for evidence, privacy and release requirements.
Vflash source is Apache-2.0. Models and adapters retain their own licenses and terms; this repository contains no model weights. The LightX2V inference framework is not a runtime dependency.
Thanks to MiniMax H3, LightX2V, and the PyTorch, Triton and CUDA communities. Sources and acknowledgements.